Benchmark OnPair Arrow GPU export

[joseph-isaacs]

Summary

• benchmark OnPair decompression directly to Arrow Device VarBin output
• retain the VarBinView benchmark as a comparison
• build and run the OnPair CUDA benchmark in the CodSpeed standalone-kernel shard

The Arrow Device path bypasses BinaryView materialization and exports device-resident validity, i32 offsets, and contiguous values buffers directly.

Local results

10 million URL-shaped strings on the local GPU:

Output	Median	Throughput
VarBinView	945.55 µs	480.52 GiB/s (516.0 GB/s)
Arrow Device VarBin	722.54 µs	628.83 GiB/s (675.2 GB/s)

Direct VarBin reduced latency by 23.6% and increased useful throughput by 30.9%.

Validation

• cargo bench -p vortex-cuda --bench onpair_cuda
• cargo test -p vortex-cuda onpair --lib (19 passed)
• cargo check -p vortex-cuda --bench onpair_cuda
• cargo +nightly fmt --all
• git diff --check

yamllint was unavailable locally.

[codspeed-hq]

Merging this PR will improve performance by 14.14%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 1 improved benchmark
✅ 1998 untouched benchmarks
🆕 2 new benchmarks
⏩ 474 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	WallTime	words_gather_scalar[65536]	9.4 µs	8.2 µs	+14.14%
🆕	WallTime	cuda/onpair/decompress_to_varbin[10M]	N/A	2.3 ms	N/A
🆕	WallTime	cuda/onpair/decompress_to_varbinview[10M]	N/A	3.6 ms	N/A

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ji/arrow-gpu-export (9b0fa11) with develop (9d28d3e)}

Footnotes

1. 474 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩